Prashant Mishra

Leading a team of 7 Data Scientists Spearheaded development of an Agentic GenAI product to dynamically select patient universes from textual input and curate treatment lines of therapy (LoT) by extracting business rules from unstructured clinical documents — reducing LoT curation from 3 months to 2 days and delivering $1.5M/year in savings for the commercial team. Led end-to-end delivery of a Multi-Agentic Insights Hub — orchestrating specialized agents (Intent, Planner, Retriever, Summarizer, Synthesis) across 15 structured and unstructured data sources — cutting insight synthesis time from 2 weeks to 1.5 hours, with wide adoption by Medical leads for strategic planning. Currently leading commercial Brand Strategic Planning (BSP) AI platform for 30+ stakeholders, aligning cross-functional interests across medical, commercial, and data teams; projected to compress the BSP cycle from 4 months to 2 weeks.

Architected and published an LLM evaluation framework incorporating Chain-of-Thought (CoT) scoring, tailored for Business Leadership consumption — adopted internally and published in the peer-reviewed PMSA Journal (Spring 2025). Delivered fast, explainable ML/DL models on high-dimensional healthcare data for Fortune 500 pharma clients including Johnson & Johnson and Astra Zeneca, translating complex outputs into actionable commercial insights. Engineered a Transformer-based language model to identify uncaptured events in patient journey data and predict future clinical events with occurrence timing, enabling precision customer targeting.

Designed a RAG-based application to extract and structure medical insights from unstructured documents into a queryable database, accelerating evidence synthesis for medical affairs teams. Developed a custom XGBoost classifier for Positive-Unlabeled (single-class) classification to impute missing patient events in claims data — awarded Highest Impact Award within the organization.

Built a Deep Learning Natural Language Generation (NLG) pipeline using a Seq2Seq Neural Machine Translation model to auto-generate textual EDA summaries from structured data. Enhanced model performance with Luong and Bahdanau Attention mechanisms, achieving a BLEU score of 31.2 using an Attentive Pointer Generator network.

Applied Statistical Process Capability Analysis and Hypothesis Testing to identify product development bottlenecks, reduce leakage costs, and improve system reliability — findings presented at the INVEST Value Engineering Conference.